Road Features Extraction Using Terrestrial Mobile Laser Scanning System

(1)

Road Features Extraction Using

Terrestrial Mobile Laser Scanning

System

by

Pankaj Kumar

A thesis presented in fulfilment of the requirements for the Degree of Doctor of Philosophy

Supervisors: Dr. Timothy McCarthy, Dr. Conor P. McElhinney

National Centre for Geocomputation Faculty of Science

National University of Ireland Maynooth Maynooth, Co. Kildare, Ireland

(2)

(3)

Declaration

This thesis has not been submitted in whole or in part to this

or any other university for any other degree and is, except

where otherwise stated, the original work of the author.

Signed:

(4)

Acknowledgments

I would like to thank my supervisor, Timothy McCarthy, for providing me this opportunity and keeping a faith from beginning to end of this Ph.D. research. A profound gratitude to my supervisor, Conor P. McElhinney, for his continuous guidance and intellectual support, without which it was diffi-cult to complete this thesis. His suggestions and positive remarks provided a new direction to my thinking towards tackling the research problems. I would thank, Paul Lewis, for taking his time out to read my thesis draft and providing valuable corrections to it. Moreover, a weekly meeting of our Mobile Mapping Group helped in terms of providing a platform, where we could have a wider discussion on various research issues.

My sincere thanks to Jan Rigby, who supported and motivated me to continue my Ph.D. research study during my difficult period, especially in the first year. To Martin Charlton, for providing me valuable suggestions and support during the thesis.

I would like to thank my all other lab mates especially Binbin, Conor C., Ishwari, Cathal, Fergal, Carson, Burchin, Ambra, Zoe, Seamus who made my stay in Ireland pleasant by creating a helpful and friendly environment. This list of acknowledgement would be incomplete without mentioning the admin-istrative and technical staff of NCG who provided me a valuable support. I would also thank to all those persons, who directly or indirectly, helped me to reach this stage of life.

I would acknowledge IRCSET Enterprise, PMS and StratAG as the re-search presented in this thesis was conducted with their financial support.

(5)

List of Figures

1.1 Road accident statistics in Europe [CAR07]. . . 12

1.2 Main causes of road accidents [TTM+₇₉_]. _{. . . .} ₁₃

1.3 Road edges: (a) kerbs in an urban road section and (b) grass-soil in a rural road section. . . 18

1.4 Road markings: (a) continuous line, triangle and (b) arrow.. . 20

1.5 Roughness along the road surface. . . 22

1.6 Quarter car simulator. . . 23

1.7 Measurement of elevation points at a 1.5 m interval along the longitudinal profile of the road surface. . . 24

1.8 Snake curve and MRI scan [LS95].. . . 28

2.1 XP-1’s digital camera image of a road section. . . 39

2.2 XP-1’s LiDAR data of a road section. . . 41

2.3 Principle of (a) multiple returns from targets and returned (b) echo pulses. . . 42

2.4 TOF method . . . 44

2.5 Phase shift method.. . . 44

2.6 Intensity and pulse width of reflected pulse. . . 45

2.7 XP-1’s navigation data along a road section. . . 46

(10)

2.9 Process for estimating navigation parameters of the mobile mapping vehicle using a GNSS base station and a navigation

system.. . . 50

2.10 Data acquisition system. . . 51

2.11 XP-1 MMS with an inset picture of the laser scanning and

navigation system mounted on the roof rack. . . 52

2.12 XP-1’s navigation data plotted over a Google earth image, highlighted in yellow, with an inset picture of its small section

(Image courtesy: Google Earth).. . . 53

2.13 Inclined laser scanner along with GPS antenna and LandINS

GPS/INS mounted on the XP-1’s roof rack. . . 54

2.14 XP-1’s (a) imaging and (b) LiDAR data of a road section. . . 56

3.1 Segmentation approaches in the road edge extraction algorithm. 59

3.2 2D Gaussian distribution with mean (0, 0) and σ = 1 [FPWW00]. 61

3.3 Convolution kernel that produces a discrete approximation to

the 2D Gaussian distribution function with mask size=5 and

σ = 1 [FPWW00].. . . 61

3.4 Hierarchical thresholding: (a) input image, (b) thresholded

object cells in the lowest resolution image, at level n and (c) thresholded neighbourhood object cells in the next lowest

res-olution image, at level n − 1 [SHB08]. . . 62

3.5 Hierarchical thresholding: (a) original LiDAR data, (b) input

slope image and (b) estimated objects. . . 63

3.6 Sobel convolution kernels, which are used to find the gradients

(11)

3.7 Canny edge detection: (a) input hierarchical thresholded

ob-jects and (b) their estimated boundaries. . . 66

3.8 Initial snake curve in the form of parametric ellipse. . . 67

3.9 Parametric ellipse snake curve initialised on a gradient image

of road object boundaries. . . 69

3.10 Traditional parametric active contour model applied to an

ob-ject with a concave boundary [XP98]. . . 70

3.11 Balloon external energy added to the snake curve. . . 71

3.12 Parametric ellipse snake curve initialised on GVF image of

road object boundaries.. . . 74

3.13 GVF active contour model applied to an object with a concave

boundary [XP98]. . . 75

3.14 Geometric active contour model. . . 76

3.15 Balloon energy pushes the snake curve and GVF energy

at-tracts the snake curve toward the object boundaries. . . 78

3.16 Final position of the snake curve. . . 79

3.17 Range dependent thresholding in the road marking extraction

algorithm. . . 79

3.18 Range dependent thresholding: (a) input 2D intensity raster

surface and (b) extracted road markings. . . 80

3.19 RANSAC in the road roughness estimation algorithm. . . 81

3.20 RANSAC surface grid fitted to the LiDAR points, with an

inset picture of a magnified portion. . . 84

4.1 Road edge extraction algorithm. . . 89

4.2 Terrain pyramids generated from the LiDAR attributes: (a)

(12)

4.3 Raster cell laid over the Voronoi polygons constituting the thinned LiDAR points in the natural neighbourhood

interpo-lation. . . 92

4.4 2D raster surfaces generated from their respective terrain

pyra-mids using natural neighbourhood interpolation: (a) slope, (b)

reflectance and (c) pulse width. . . 93

4.5 Snake curve initialisation in (a) parametric ellipse form and

(b) centre of the road estimation. . . 95

4.6 φ angle is calculated from θ, which is an average heading angle

of the mobile van along the road section under investigation

that can lie in between (a) 0◦ and 90◦, (b) 90◦ and 180◦, (c)

180◦ and 270◦ and (d) 270◦ and 360◦. . . 96

4.7 Snake curve positions: (a) initial, (b) iterative and (c) final. . 97

4.8 Intersection point in between two overlapped snake curves. . . 98

4.9 Three processed road sections with (a) first intersection points

highlighted, (b) second intersection points highlighted and (c)

final left and right road edges. . . 100

4.10 3D left and right road edges. . . 101

4.11 Road edge validation algorithm. . . 102

4.12 Slope surface estimated from the elevation raster surface gen-erated using linear interpolation applied to (a) full resolution, (b) first level terrain pyramid and using natural neighbour-hood interpolation applied to (a) full resolution, (b) first level

(13)

4.13 Reflectance raster surface generated from (a) full resolution terrain pyramid using linear interpolation and (b) first level terrain pyramid using natural neighbourhood interpolation. Pulse width raster surface generated from (c) full resolution terrain pyramid using linear interpolation and (b) first level

terrain pyramid using natural neighbourhood interpolation. . . 106

4.14 Final position of the snake curve over the slope raster surface

with (a) 0.02 m2, (b) 0.04 m2, (c) 0.06 m2, (d) 0.08 m2, (e) 0.1

m2 _{and (f) 0.2 m}2 _{cell size.} _{. . . 109}

4.15 Plot of completeness obtained from the snake curve in the six

test cases of an optimal cell size estimation analysis.. . . 110

4.16 Plot of the time taken by the snake curve to move from its initial position to the final position in the six test cases of an

optimal cell size estimation analysis. . . 111

4.17 3D LiDAR data: (a) 6 m ×10 m, (b) 4.5 m ×5 m and (c) 3 m

×4 m. . . 112

4.18 LiDAR point data along the (a) left and (b) right side of the

road section. . . 113

4.19 Road edge digitised from the LiDAR data: (a) 3.2 m ×3.3 m

and (b) 1.8 m ×2 m. . . 113

4.20 Box plots for the (a) 2D and (b) 3D accuracy of the extracted left edges in the six test cases of an optimal cell size estimation

analysis. . . 114

4.21 Box plot for the (a) 2D and (b) 3D accuracy of the extracted right edges in the six test cases of an optimal cell size

(14)

4.22 Missed points as circled in blue in the section of LiDAR data. 120

with (a) 10 m, (b) 20 m and (c) 30 m road length. . . 121

4.24 Final position of the snake curve over the slope raster surface in the (a) first, (b) second, (c) third, (d) fourth and (e) fifth

case of the internal energy parameter analysis. . . 123

4.25 Final position of the snake curve over the slope raster surface in the (a) first, (b) second, (c) third, (d) fourth, (e) fifth, (f) sixth, (g) seventh, (h) eighth and (i) ninth case of the external

energy parameter analysis. . . 126

4.26 Hierarchical thresholding applied to the reflectance raster sur-face provided the objects with road marking cells near the

road edge points as circled in blue. . . 127

4.27 Hierarchical thresholding applied to the slope raster surface was not able to remove the noisy cells in between the road

edge points. . . 129

with (a) κ4 = 1 and (b) κ4 = 3 parameters. . . 130

4.29 Digital image of the rural road section consisting of grass-soil

edges (Geographic location: 53◦34028.0700N 7◦10013.7600W). . . 131

4.30 Final position of the snake curve over the slope raster surface obtained through the manual selection of parameters in the (a) first, (b) second, (c) third, (d) fourth, (e) fifth and (f)

sixth data section in the rural road. . . 133

4.31 Extracted 3D left and right edges in the rural road section

(15)

4.32 Digital image of the urban road section consisting of kerb edges

(Geographic location: 53◦36037.4200N 7◦5041.0000W). . . 135

sixth data section in the urban road. . . 136

4.34 Extracted 3D left and right edges in the urban road section

obtained through the manual selection of parameters. . . 137

4.35 Digital image of the national primary road section consist-ing of grass-soil edges with shoulders (Geographic location:

53◦3808.8000N 7◦29011.0600W). . . 138

sixth data section in the national primary road. . . 139

4.37 Extracted 3D left and right edges in the national primary road

section obtained through the manual selection of parameters. . 140

4.38 Automated final position of the snake curve over the slope raster surface for the (a) first, (b) second, (c) third, (d) fourth,

(e) fifth and (f) sixth data section in the rural road. . . 142

4.39 Automatically extracted 3D left and right edges in the rural

4.40 Automated final position of the snake curve over the slope raster surface for the (a) first, (b) second, (c) third, (d) fourth,

(16)

4.41 Automatically extracted 3D left and right edges in the urban

4.42 Automatic final position of the snake curve for the (a) first, (b) second, (c) third, (d) fourth, (e) fifth and (f) sixth data

section in the national primary road. . . 146

4.43 Automatically extracted 3D left and right edges in the national

primary road section. . . 147

4.44 Box plot for the (a) 2D and (b) 3D accuracy of the manually and the automatically extracted left edges in the rural road

section.. . . 149

4.45 Box plot for the (a) 2D and (b) 3D accuracy of the manually and the automatically extracted right edges in the rural road

section.. . . 150

4.46 Box plot for the (a) 2D and (b) 3D accuracy of the manually and the automatically extracted left edges in the urban road

section.. . . 154

4.47 Box plot for the (a) 2D and (b) 3D accuracy of the manually and the automatically extracted right edges in the urban road

section.. . . 155

4.48 Box plot for the (a) 2D and (b) 3D accuracy of the manu-ally and the automaticmanu-ally extracted left edges in the national

4.49 Box plot for the (a) 2D and (b) 3D accuracy of the manually and the automatically extracted right edges in the national

(17)

5.2 Snake curve is (a) laid over the LiDAR points to (b) estimate

the road surface. . . 170

5.3 2D raster surfaces generated from the LiDAR data: (a)

inten-sity and (b) range. . . 171

5.4 Navigation data is used to select a range value to apply

mul-tiple threshold values to the intensity raster surface. . . 172

5.5 Side view of the non-planar road surface. . . 173

5.6 Road markings extracted from the intensity raster surface

us-ing range dependent thresholdus-ing. . . 174

5.7 Structuring elements: (a) diamond shaped with radius = 1,

(b) linear shaped with length = 3 and angle, φ0 = 45◦ and (c)

linear shaped with length = 5 and angle, φ0 = 90◦.. . . 174

5.8 Estimation of φ0 angle of the linear shaped structuring element

from the average heading angle of the mobile van, θ, along the

road section that can lie in between (a) 0◦ and 90◦, (b) 90◦

and 180◦, (c) 180◦ and 270◦ and (d) 270◦ and 360◦. . . 176

5.9 Dilation operation: (a) input binary image with an inset

pic-ture of their road marking cells and (b) dilated image with an

inset picture of their dilated road marking cells. . . 177

5.10 Noise removal process: (a) input dilated image and (b) noise

cells removed from it. . . 178

5.11 Erosion operation: (a) input dilated image with an inset pic-ture of their road marking cells and (b) eroded image with an

inset picture of their eroded road marking cells. . . 179

5.12 Extracted road markings (a) before and (b) after applying the

(18)

5.13 LiDAR points (a) belonging to the road surface and the

ex-tracted (b) 3D road markings. . . 180

5.14 Road markings extracted from the intensity image with (a)

0.01 m2_{, (b) 0.04 m}2_{, (c) 0.06 m}2_{, (d) 0.08 m}2 _{and (e) 0.1 m}2

cell size. . . 182

5.15 Range dependent thresholding applied to the intensity raster surface in the road section with (a) narrower and (b) greater

width. . . 185

5.16 Road markings extracted using the (a) lower, (b) higher and (c) optimal threshold value applied to the intensity raster

sur-face. . . 186

5.17 Digital image of (a) rural and (b) urban road section contain-ing both broken and continuous line markcontain-ings (Geographic

lo-cations: (a) 53◦34028.0700N 7◦10013.7600W and (b) 53◦36033.4300N

7◦5046.9600W). . . 187

5.18 2D broken and continuous line markings extracted from the (a) first, (b) second, (c) third, (d) fourth, (e) fifth and (f) sixth

data section of the rural road. . . 189

5.19 2D broken and continuous line markings extracted from the (a) first, (b) second, (c) third, (d) fourth, (e) fifth and (f) sixth

data section of the urban road. . . 190

5.20 Road markings extraction: (a) original LiDAR data and (b) extracted 3D broken and continuous line markings of the rural

(19)

5.21 Road markings extraction: (a) original LiDAR data and (b) extracted 3D broken and continuous line markings of the

ur-ban road section. . . 192

5.22 Digital image of the rural road section containing word mark-ings along with broken and continuous line markmark-ings

(Geo-graphic location: 53◦34050.54600N 7◦8057.6200W). . . 193

5.23 Output snake curve in the rural road section containing word

markings along with broken and continuous line markings. . . 194

5.24 Road markings extraction: (a) original LiDAR data and (b) extracted 3D word, broken and continuous line markings of

the rural road section. . . 194

5.25 Digital image of the national primary road section containing

hatch and broken line markings (Geographic location: 53◦33049.87800N

7◦21024.14800W). . . 195

5.26 Output snake curve in the national primary road section

con-taining hatch and broken line markings. . . 196

5.27 Road markings extraction: (a) original LiDAR data and (b) extracted 3D hatch and broken line markings of the national

5.28 Digital image of the national primary road section containing

arrow and broken line markings (Geographic location: 53◦35030.55800N

7◦22028.42800W). . . 198

5.29 Output snake curve in the national primary road section

(20)

5.30 Road markings extraction: (a) original LiDAR data and (b) extracted 3D arrow and broken line markings of the national

5.31 Digital image of the urban road section containing pedestrian crossing, broken transverse and zig-zag markings (Geographic

location: 53◦36043.72100N 7◦5032.87100W). . . 200

5.32 Output snake curve in the urban road section containing

pedes-trian crossing, broken transverse and zig-zag markings. . . 201

5.33 Road markings extraction: (a) original LiDAR data and (b) extracted 3D pedestrian crossing, broken transverse line and

zig-zag markings of the urban road section. . . 201

5.34 Digital image of the urban road section containing zig-zag

markings (Geographic location: 53◦39019.22500N 7◦3109.79200W).202

5.35 Output snake curve in the urban road section containing

zig-zag markings. . . 203

5.36 Road markings extraction: (a) original LiDAR data and (b)

extracted 3D zig-zag markings of the urban road section. . . . 203

5.37 2D hatch and broken line markings of the national primary road section with the noise (a) removed only once before ap-plying the erosion operation and (b) removed twice, before

and after applying the erosion operation. . . 206

6.1 Road roughness estimation algorithm.. . . 210

6.2 Snake curve is (a) laid over the LiDAR data and (b) the points

(21)

6.3 Input LiDAR and navigation data in the (a) 3D, (b) 2D plane and rotated LiDAR and navigation data in the (c) 3D, (d) 2D

plane. . . 212

6.4 Surface grid is fitted to the LiDAR points that belong to (a)

whole road surface and (b) the left side of the road surface. . . 213

6.5 RANSAC surface grid fitted to the LiDAR points in the (c)

3D and (d) 2D plane. . . 214

6.6 Inversely rotated surface grid and LiDAR points along the

navigation track of the mobile van in the (c) 3D and (d) 2D

plane. . . 215

6.7 Digital image of (a) first, (b) second and (c) third section of

ur-ban road (Geographic locations: (a) 53◦36033.6800N 7◦5046.3900W,

(b) 53◦36036.58700N 7◦5041.67100W and (c) 53◦36040.58400N 7◦5037.78900W).216

6.8 Output snake curve in the (a) second and (b) third section of

urban road. . . 217

6.9 Duplicate points were removed in the first LiDAR section to

provide non-overlapped sections. . . 218

6.10 Surface grids and LiDAR points along the navigation track in

the (a) first, (b) second and (c) third section of urban road.. . 220

6.11 Digital image of the national primary road section (Geographic

locations: 53◦38014.40700N 7◦29024.62200W). . . 221

6.12 Inversely rotated surface grids and LiDAR points along the

navigation track in the national primary road section. . . 222

6.13 Road surface deviation maps for the first urban road section:

(22)

6.14 Road surface deviation maps for the second urban road

sec-tion: (a) 3D and (b) 2D. . . 224

6.15 Road surface deviation maps for the third urban road section:

(a) 3D and (b) 2D. . . 225

6.16 Road surface deviation maps for the national primary road

section: (a) 3D and (b) 2D. . . 226

6.17 Plot of the LiDAR points and the surface grid points along the navigation track in (a) the first and (b) the second urban

6.18 Plot of the LiDAR points and the surface grid points along the navigation track in (a) the third urban and (b) the national

6.19 Box plot of the standard deviation of the elevation residual points along the navigation track in (a) the first and (b) the

second urban road section. . . 229

6.20 Box plot of the standard deviation of the elevation residual points along the navigation track in (a) the third urban and

(b) the national primary road section.. . . 230

(23)

List of Tables

1.1 Examples of terrestrial MMSs. . . 6

2.1 IXSEA LandINS navigation system specifications. . . 53

2.2 Riegl VQ-250 specifications. . . 55

4.1 Global minimum and maximum values of the LiDAR attributes.107

4.2 Different parameters used in the six test cases of an optimal

cell size estimation analysis. . . 108

4.3 Statistical analysis of the 2D accuracy of the left edges along

with completeness and time in the six test cases. . . 116

4.4 Statistical analysis of the 3D accuracy of the left edges along

4.5 Statistical analysis of the 2D accuracy of the right edges along

4.6 Statistical analysis of the 3D accuracy of the right edges along

4.7 Different parameters used in the three test cases of an optimal

road length estimation analysis. . . 120

4.8 Internal energy weight parameters used in the five test cases. . 122

(24)

4.10 LiDAR point density over the left and right sides of different

road sections, acquired with the XP-1 system. . . 128

4.11 φ angle calculated from θ, average heading angle in each

nav-igation section in the rural road. . . 132

nav-igation section in the urban road. . . 135

nav-igation section in the national primary road. . . 138

4.14 Hierarchical threshold parameters selected empirically for each

4.15 Statistical analysis of the 2D and 3D accuracy of the manually and the automatically extracted left edges in the rural road

section.. . . 151

4.16 Statistical analysis of the 2D and 3D accuracy of the manually and the automatically extracted right edges in the rural road

section.. . . 151

4.17 Statistical analysis of the 2D and 3D accuracy of the manually and the automatically extracted left edges in the urban road

section.. . . 153

4.18 Statistical analysis of the 2D and 3D accuracy of the manually and the automatically extracted right edges in the urban road

section.. . . 156

4.19 Statistical analysis of the 2D and 3D accuracy of the manu-ally and the automaticmanu-ally extracted left edges in the national

(25)

4.20 Statistical analysis of the 2D and 3D accuracy of the manu-ally and the automaticmanu-ally extracted left edges in the national

5.1 Maximum and minimum values of the LiDAR intensity and

range attributes. . . 181

5.2 Length and average width values of the extracted road

mark-ings in the five cases. . . 183

5.3 φ0 angle calculated from θ angle in each navigation section of

rural road section.. . . 188

5.4 φ0 angle calculated from θ angle in each navigation section of

urban road section. . . 188

6.1 φ angle calculated from θ, average heading angle in the

navi-gation sections of second and third urban road section. . . 217

6.2 t and parameters used for fitting the surface grid to each

LiDAR section of the urban road. . . 219

6.3 t and parameters used for fitting the surface grid to each

LiDAR section of the national primary road. . . 219

6.4 Statistical analysis of the standard deviation of the elevation

residual points along the navigation track in the first, second,

(26)

Abstract

In this thesis, we present the experimental research and key contributions we have made in the field of road feature extraction from LiDAR data. We de-tail the development of three automated algorithms for the extraction of road features from terrestrial mobile LiDAR data. LiDAR data is a rich source of 3D geo-referenced information whose volume and scale have inhibited the development of automated algorithms. Automated feature extraction algo-rithms enable the wider geospatial industry to transition from traditional road feature surveying approaches to terrestrial mobile laser scanning tech-nologies.

Our first contribution to this field is an automated road edge extraction algorithm which can be applied to LiDAR data and navigation information acquired by mobile survey vehicles. This novel algorithm relies on the com-bination of thresholding and a parametric active contour model to precisely extract road edges. We describe an automated validation algorithm we de-veloped to determine the accuracy of our road edge extraction algorithm.

Using the extracted road edges, we are able to accurately extract the road surface from the LiDAR data. This enables us to develop an efficient automated road marking extraction algorithm which is our second contribu-tion. Through the thresholding of the intensity values of road surface LiDAR points, we can extract the road marking LiDAR points. The third contri-bution of this thesis is the development of an automated road roughness estimation algorithm which is also dependent on the accurate detection of road surface LiDAR points. We fit a surface grid to the LiDAR points rep-resenting an ideal road surface and measure the elevation difference between this surface and the actual LiDAR points to compute the surface deviation

(27)

along a track representing a vehicle wheel.

We automated these algorithms through exhaustive examination of opti-mal parameters and methods for their implementation. To verify these novel algorithms, we tested them on varying types of road sections representing rural, urban and national primary road sections. The research work carried out in the course of this thesis provides valuable insights as well as a pro-totype road feature extraction tool-set, for both national road authorities and survey companies. These findings and knowledge contribute to a more rapid, cost-effective and comprehensive approach to surveying road networks which, in turn, enables a more efficient, comfortable and safer journey for all road users.

(28)

Acronyms

C/A Coarse Acquisition

CCD Charge Coupled Device

CDSS Car Driven Survey System

CMOS Complementary Metal Oxide Semiconductor

CW Continuous Wave

DGPS Differential Global Positioning System

DIA Direct Inertial Aiding

DLM Digital Landscape Model

DMI Distance Measuring Instrument

DSM Digital Surface Model

DTM Digital Terrain Model

EU European Union

FOGs Fibre Optic Gyroscopes

FOV Field Of View

GCPs Ground Control Points

GDP Gross Domestic Product

GIS Geographic Information System

GNP Gross National Product

GNSS Global Navigation Satellite System

GVF Gradient Vector Flow

IMC Immersive Media Coorporation

IMU Inertial Measurement Unit

INS Inertial Navigation System

(29)

IRI International Roughness Index

KISS KInematic Surveying System

LAN Local Area Network

LED Light Emitting Diode

LiDAR Light Detection And Ranging

LSB Least Square B-spline

MRI Magnetic Resonance Imaging

MMS Mobile Mapping System

NDSM Normalised Digital Surface Model

NDVI Normalised Differential Vegetation Index

NRA National Roads Authority

NSM Network Safety Management

NUIM National University of Ireland Maynooth

PPK Post Processing Kinematic

PPS Pulse Per Second

QCS Quarter Car Simulator

RANSAC RANdom SAmple Consensus

RINEX Receiver INdependent EXchange

RLG Ring Laser Gyro

RMS Root Mean Square

RSA Road Safety Audit

RSI Road Safety Inspection

RTRRMs Response Type Road Roughness Meters

RTK Real Time Kinematic

TIN Triangulated Irregular Network

(30)

UHF Ultra High Frequency

VISAT Video Inertial SATellite

WHO World Health Organization

(31)

Chapter 1 Introduction

1

The demand for accurate 3D mapping of natural environments and man-made features has increased due to the spatial detail required by scientists, engineers and planners [BMS06]. Light Detection And Ranging (LiDAR) is a relatively recent technology, enabling 3D modelling of real world environment by measuring the time of return of an emitted light pulses. The information obtained through laser scanning systems, which use LiDAR technology, have application in road safety, urban planning, flood plain, glacier and avalanche mapping, bathymetry, geomorphology, forest survey, bridge and transmission line detection [Bur02]. Laser scanning systems enable the acquisition of an accurately georeferenced set of dense LiDAR point cloud data [PT08]. Other benefit of this type of system is the high level of automation during data capture and the ability of this system to acquire data beneath tree’s canopy. Laser scanning systems are used to acquire LiDAR data from aerial and terrestrial platforms. The data acquired from these systems differs in terms of its intrinsic accuracy and resolution for a variety of reasons but primarily

1_{Throughout the thesis, the terms ’we’ and ’our’ are used to describe the doctorate}

(32)

due to the distance of the scanner to the target objects [ROPV09]. In recent years, the use of laser scanners onboard terrestrial based moving vehicles has increased for the collection of high quality 3D data. The applicability of these terrestrial mobile laser scanning systems continue to prove their worth in route corridor mapping due to the rapid, continuous and cost effective 3D data acquisition capability compared with static terrestrial laser scanning systems [HPKH08, BMS06]. LiDAR data records a number of attributes including elevation, intensity, pulse width, range and multiple echo informa-tion, all of which can be used for extracting road features. The volume of data produced by a terrestrial mobile laser scanning system such as Riegl VQ-250 is large, generating 300,000 points per second resulting in approximately 20 GB of data per hour. However, manual processing of LiDAR data for road features extraction is very time consuming. These sensor characteristics, po-tential applications and challenges provide the underlying motivation for this thesis.

We present three key contributions we have made to the field of auto-mated road feature extraction from LiDAR data. Correct identification of the road boundaries is essential in order to obtain a precise estimation of road geometry and associated features. Our first contribution is the development of an automated algorithm for extracting road edges from terrestrial mo-bile LiDAR data. A priori knowledge of the road boundaries and associated surface in the LiDAR data facilitates efficient extraction of the road mark-ings and roughness. The second contribution deals with the development of an automated algorithm for extracting road markings from terrestrial mo-bile LiDAR data. The third contribution focuses on the development of an automated algorithm for estimating road roughness from terrestrial mobile

(33)

LiDAR data.

In this chapter, we introduce the reader to terrestrial Mobile Mapping Systems (MMSs). In Section 1.1, we describe terrestrial MMSs and their ability to provide georeferenced spatial data. We review various MMSs com-prising digital cameras, navigation sensors and laser scanners over the past two decades. In Section 1.2, the importance of road safety and its linkage to road geometry and road features is discussed. Terrestrial MMSs can be used to acquire 3D information about the road environment that can, in turn, as-sist decision makers in identifying safety risk elements along road networks. We discuss an application of terrestrial MMSs in road safety. In Section

1.3, we review various methods developed for extracting road features from LiDAR data. We investigate methods based on snake curves that has been developed for segmentation. Following the review, we identify the research limitations which have been addressed by doctorate research. Section 1.4

deals with the key contributions of this thesis to the field of road feature extraction from LiDAR data. Finally, an outline of the thesis is presented in Section 1.5.

1.1 Mobile Mapping

Mobile mapping refers to a means of collecting geospatial data using mapping and navigation sensors that are mounted rigidly onboard a mobile platform [TL07]. The concept of mobile mapping dates back to early 1990’s and since then has been primarily driven by the advances in kinematic positioning, machine vision, laser scanning systems, data fusion and spatial information technologies. The mobile platform term relates to the transportation mode of the MMS which can be land-based (car, train), air-borne (aircraft) or

(34)

marine (ship, submarine) [Nov93]. Mapping sensors can consist of imaging and laser scanning system while the navigation system is based on integrated Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS). The navigation system may also be complemented with dead reckoning sensors such as a Distance Measuring Instrument (DMI), an odometer or a digital compass.

The effectiveness of mobile mapping lies in its ability to directly georef-erence the acquired spatial and spectral data with the recorded navigation data within the global coordinate frame. The direct georeferencing provided by MMSs removes the need for Ground Control Points (GCPs) or any other external geographic referencing system. This is different to some other spa-tial data acquisition systems where a suitable number of well spread GCPs are required from a field survey prior to or after data acquisition [EES02]. However, GCPs can be used to increase the data accuracy in MMSs where higher levels of accuracy are required for some projects. In the following sections, we review various terrestrial MMSs developed using digital cameras and laser scanners.

1.1.1 Mobile Mapping with Digital Cameras

Terrestrial MMSs have been actively researched and developed over the past two decades [BMS06]. In their initial development phase, the MMSs were developed based on GNSS/INS integration and digital cameras which were used for the acquisition of road inventories and surrounding route corridor environment. Some examples of terrestrial MMSs with digital cameras have been listed in Table 1.1. The first terrestrial MMS, GPSVan, was developed by the Centre of Mapping at The Ohio State University in the early 1990’s

(35)

Name & Dev elop ers Na vigation Sen s ors Mapping Sensors P osition A ccuracy Primary Application GPSV an, Ohio State Univ ersit y GPS, gyro based INS, o dometer Digital stereo-vision system, colour video cameras 1 to 3 m Digital road maps and high w a y in v en tori e s VISA T, Univ ersit y of Calgary DGPS, strap do wn INS, DMI eigh t digital came ras , one colour video cam-era 0.3 m High w a y surv eying sys-tem KISS, Univ e rs it y of th e F ederal Armed F orces Munic h GPS, IMU, tw o in-clinometers, o dome ter, barometer P air of digital cam-eras, one colour video camera 0.1 m Road features an d its nearb y surrounding mapping CDSS, Geo detic Insti-tute Aac hen tw o C/A co de GPS, tw o o dometers, barometer tw o video cameras Less than 1 m in high-w a ys and up to 3 m in cities Road features mapping using lo w cost hardw are system PHOTOBUS, Swiss F ederal Institute of T ec hnology Dual frequency GPS, INS V ertical and hori -zon tal orien ted video cameras 0.2 to 0.4 m Road signs mapping T ruc kMap, John E. Chance & Asso ciates Multiple GPS re ceiv ers, digital attitude sensor Laser range finder, video camera 0.31 m in horizon tal Electric transmission line corridor surv eying, urban mapping

(36)

GEOMOBIL, Carto-graphic Institute of Catalonia tw o sets of GPS, IMU, DMI P air of stereoscopic digital cameras, laser scanner 0.18 m and 0.35 m in horizon tal while 0.13 m in v ertical p os i-tions Urban mapping StreetMapp er, 3D Laser Mapping and IGI mBH GPS, IMU, DIA Digital camera, tw o laser scanners 0.01 to 0.03 m High w a ys, road features and urban m ap ping IP-S2, T op con GNSS, IMU, o dometer P anoramic camera unit, three laser scanners 0.015 to 0.040 m Utilities, GIS asset managemen t an d trans-p ortation L YNX M1, Optec h GPS, IMU, DM I four digital cameras, tw o laser scanners 0.05 m Road infrastructure, as-set managemen t, utili-ties MoSES, 3D Mapping Solutions Gm bH DGPS, IMU, o dometer eigh t digital came ras , tw o laser scanners 0.03 m in horizon tal and 0.003 m in v erti-cal p osition Road and rail construc-tion site surv eys, tra-jectory qualit y manage-men t, urb a n mapping StreetView, Go ogle DGPS, IMU, o dometer nine digital camera, three laser scanners N.A. Street mapping for visu-alisation purp ose T able 1.1: Examples of terrestrial MMSs.

(37)

[Goa91]. The absolute object accuracy was 1-3 m in Easting and Northing which describes the accuracy achieved by GPSVan in object positioning with respect to the ground truth values. This limited accuracy was attributed to the use of a code-only Global Positioning System (GPS) receiver and gyro based inertial system [EES02]. In 1993, the Video Inertial SATellite (VISAT) system was developed at the University of Calgary. This system was based on dual frequency carrier-phase Differential Global Positioning System (DGPS) and a more accurate Inertial Measurement Unit (IMU) [SMES+93]. The VISAT system reported an improved horizontal accuracy of 0.3 m in Easting and Northing while operating at a speed of up to 60 km/hr but it had drawbacks including increased cost and a much higher level of system complexity [Es05].

By the mid 1990’s additional systems based on a similar architecture were developed worldwide. In 1995, the KInematic Surveying System (KISS) was designed and developed by the University of the Federal Armed Forces for kinematic Geographical Information System (GIS) data acquisition [HCH+₉₅_]. During post-mission mode, the surveyed data was processed to produce 3D georeferenced information of the road network and its surrounding environ-ment. In 1998, the Car Driven Survey System (CDSS) system was developed at the Geodetic Institute Aachen for acquiring road data using a low-cost navigation system which comprised a Coarse Acquisition (C/A) code GPS, odometer and barometer [BA98]. In 2000, the Geodetic Engineering Labo-ratory of the Swiss Federal Institute of Technology, Lausanne designed the PHOTOBUS system which was distinguishable from other systems by its ability to georeference the road centre line using a camera oriented vertically and acquire road sign data using a camera oriented horizontally [GSG03].

(38)

The object accuracy achieved was 20-40 cm with respect to the central axis of the road.

1.1.2 Mobile Mapping with Laser Scanners

Laser scanning has been accepted as an effective method for spatial data acquisition in MMSs due to the highly accurate and dense point cloud data that can be captured and recorded by these systems. The integration of laser scanners with terrestrial MMSs facilitated the rapid and cost effective captur-ing of 3D data for larger urban areas [HPKH08]. Some examples of terrestrial MMSs with laser scanners have also been listed in 1.1. In 1995, the applica-bility of terrestrial mobile laser scanning system for surveying and mapping transmission line corridors was demonstrated by Reed et al. [RLW96]. Their van-based system, TruckMAP, incorporated real time positioning, a reflector-less laser range finder and high resolution video cameras which were used to map the sub-stations and electric transmission line corridor while the sections of transmission line with no vehicular access were surveyed with an airborne laser scanning system, FLI-MAP. In 2000, the Cartographic Institute of Cat-alonia developed the GEOMOBIL system which included all the navigation and mapping sensors required for acquiring digital stereo pair images and subsequent direct georeferencing [TBA+₀₄_{]. Later, the group integrated a} laser scanner into their system which was able to collect 10,000 points per second [TAB+₀₄_{]. The absolute accuracy of the laser scanning system was} measured as 0.18 m in Easting, 0.35 m in Northing and 0.13 m in the vertical plane. The design and development of several other MMSs based on laser scanners were reported over subsequent years [ZS03, GNA+06, KAS+07].

(39)

terres-trial MMSs to be broadened to include 3D route corridor and urban map-ping, traffic simulation studies, virtual reality modelling and utility mapping. This expansion of application areas has led to the development of a number of commercial enterprises based around MMS across the world. Over the last 20 years, MMSs have slowly developed from research projects in the aca-demic sector to becoming commercially viable activities [Pet10]. According to a market research study conducted by ARC advisory group, the 3D laser scanning market is expected to double in size from 2010 to 2015 [Rio11]. There are a number of companies which provide MMS products for fast and automated data acquisition. 3D Laser Mapping is one such company which developed the StreetMapper system in collaboration with IGI mbH to meet the requirements of clients for rapid 3D mapping of highways, road fea-tures, buildings and infrastructure using vehicle-mounted lasers [HC10]. The StreetMapper system, which has been operating since early 2005, utilises a GPS receiver, fibre optic gyro based IMU, Direct Inertial Aiding (DIA) sys-tem, high resolution digital camera and two Riegl VQ-250 laser scanners. Each scanner offers a 3600 Field Of View (FOV), a range of 300 m and a measurement rate of 300,000 points per second. In 2009, another company, Topcon positioning systems introduced IP-S2 (Integrated Positioning) MMS which integrates a Topcon dual frequency GNSS receiver operating at 20 Hz, Honeywell HG1700 tactical grade IMU based on Ring Laser Gyro (RLG), wheel-mounted odometer, the Ladybug multi-camera unit that is capable of 3600 panoramic imaging and three LMS 291 laser scanners [Top10]. One other company, Optech, provides a LYNX mobile mapping product which incorporates an Applanix POSPac navigation system, imaging system and laser scanners that are built in-house by Optech [Opt10]. Their LYNX M1

(40)

model, launched in 2010, provides a laser measurement rate of 500,000 points per second, scan frequency of 200 Hz and range of 200 m. Some of the other MMS product suppliers include Mitsubishi Electric Corporation, Trimble and Riegl which are all well-established in the mapping and surveying industry [Pet10].

Apart from these MMS product suppliers, there are numerous companies which offer mapping services. 3D Mapping Solutions GmbH offers services in kinematic surveying of road networks with their mobile road mapping system MoSES. Their system integrates DGPS, IMU, linear odometer, eight multispectral cameras and two laser scanners [Gra08]. Google provides Street View images to the users online which are acquired using their MMS Street View cars. Their vehicle comprise a Topcon DGPS/IMU positioning system, a wheel mounted odometer, three SICK LMS 291 laser scanners and nine Elphel digital cameras that are configured to provide 3600_{horizontal and 290}0

degree vertical panoramic view [Wil07, Pet10]. Most recently, Google has introduced a new tricycle platform, named Trikes which are equipped with a similar set of positioning system, laser scanners and digital cameras. These Trikes are being used for data collection in areas which are not accessible by the cars. Similarly, other commercial companies like Tele Atlas and NAVTEQ provide digital map databases for navigation and cartographic applications which they generate using their own mobile mapping vehicles equipped with positioning and imaging systems [Pet10]. MMSs have many applications and in the next section, we describe one area where they can be used to improve safety along road networks.

(41)

1.2 Road Safety Applications of Mobile

Map-ping

Road transportation plays a vital role in the progress and socio-economic growth of society enabling the safe movement of goods, people and services. Roads are designed and built based on numerous design criteria, notably, travel time, user comfort and convenience, fuel consumption, construction, cost and environmental impact [ETS97]. A well designed and maintained route infrastructure assists in driver safety as well as in the efficient use of overall network in terms of route navigation.

Road accidents have become one of the main concerns for policy mak-ers and road infrastructure developmak-ers due to thousands of deaths and the economic loss caused by them. Each year, around 1.2 million people die in road crashes around the world while around 50 million are severely injured [WHO09]. Furthermore, these accidents cost between 1 and 2 % of a coun-try’s annual Gross National Product (GNP) [WHO10]. According to the World Health Organization (WHO) report, road traffic accidents are likely to become the fifth leading cause of death in the world by 2030 [WHO11]. In the member states of the European Union (EU), road traffic accidents claim around 35,000 lives and leave more than 1.6 million people injured annu-ally [BEY+₁₀_{]. The economic cost has been estimated at around 2% of EU} countries Gross Domestic Product (GDP), around 180 billion Euro, which is twice the EU’s annual budget [Saf09]. The statistics of road accidents, fatalities and injuries occurred in the EU member states from 1990 − 2006 is graphically represented in Figure1.1. The negative impact of road accidents can not be ignored in terms of the very sizeable social and economic loss.

(42)

Figure 1.1: Road accident statistics in Europe [CAR07].

Thus, the main challenge for policy makers is to ensure that road networks are as safe as possible whilst maintaining quality and mobility.

The main causes of road accidents can be attributed to driver-behaviour, vehicle and road infrastructure or a combination of all these as described in Figure1.2[TTM+79]. Although driver behaviour is the main cause, the other two factors, vehicle and the road infrastructure, usually contribute to the final outcome. Road transport networks should be developed and maintained by taking into account the interaction between the above mentioned three factors. To date less consideration has been given to the road infrastructure element [IRF03]. Analysis shows that accidents occur due to human error mostly at specific accident hotspots. Road design has an immediate effect on accident risk as it influences driver behaviour in terms of speed, acceleration and lateral position. Safe road-way infrastructure has an important role in reducing the accident risk as road infrastructure contributes to one out of three fatal accidents [UNE08]. Road safety considerations must result in a

(43)

Figure 1.2: Main causes of road accidents [TTM+79].

road environment that should be self-explaining and forgiving, in the sense that users are not faced with unexpected situations and their mistakes can be, if not avoided, corrected [ERS06].

Recent research investigations have described a significant correlation be-tween road infrastructure and accident analysis values [GPG+₀₇_]. _Road user safety may be affected by road geometry and physical factors along the route corridor. Road geometry includes the parameters used for designing roads such as horizontal length section, curve radius, curvature change radius (CCR), vertical grade, cross-sectional lane width, shoulder width, median, number of lanes and stopping sight distance [GPG+₀₇_{]. The physical factors} refer to the objects along the route corridor such as traffic signs, light poles, trees, walls and signage. Route safety also depends on the existence and con-dition of road safety interventions along the roads. For example, road signs may be missing or suffer from reduced visibility due to temporary occlusion arising from vegetation growth, weather or from some other factors. Road geometry and physical road factors are required to be located, measured,

(44)

classified and recorded in a timely, cost effective manner in order to schedule maintenance and ensure maximum safety conditions for road users.

Various safety schemes and standards such as Road Safety Audit (RSA), Road Safety Inspection (RSI) and Network Safety Management (NSM) are implemented to qualitatively estimate potential road safety issues along the route corridor. The aim of these safety schemes is to identify the elements of the road that may present a safety concern and explore the various opportuni-ties to eliminate identified safety concerns [ETS97]. Current road surveys col-lect this information manually which usually involves an engineer annotating a digital map or using spatially referenced video to manually classify various features along the route [ERE+₀₈_{]. The information collected through these} surveys is sometimes incomplete and insufficient for qualitative estimation of potential road safety issues. It can also be time consuming and expensive to conduct these inspections on a large scale. A recent research call highlighted the requirement for common evaluation tools and implementation strategies in carrying out these inspections and assessing risk along route corridors [PME+09]. One research project EuRSI [MM10] demonstrated that MMS could be used to collect physical route corridor information for rapid safety analysis.

With the potential of GIS technologies in road management, terrestrial MMSs present a reliable and cost effective alternative for carrying out road inspections along the route corridor. Terrestrial MMSs can be employed to capture 3D spatially referenced information about road geometry and physical road objects. This information can assist decision makers to identify the possible risk elements of the road which may present a safety concern. In the next section, we review various methods developed for extracting road

(45)

features from LiDAR data

1.3 Road Features Extraction

Accurate information about the road and its features is a prerequisite for effective management of road networks and to ensure maximum safe driv-ing condition for road users. The extraction of road networks from aerial and satellite multi spectral optical images has been extensively researched. However, some limiting factors such as shadows, complex illumination and spatial accuracy prevailed in those approaches [SNG10]. The use of LiDAR technology for mapping road infrastructure provides accurate and dense 3D point cloud data which contain elevation, intensity, pulse width, range and multiple echo information. These data attributes can be used for reliable and precise extraction of the different road features. The methods developed for segmenting LiDAR data are mostly based on the identification of planar or smooth surfaces and the classification of point cloud data based on its attributes [Vos09]. In a related area, several methods have been developed over the past decade for extracting urban building features from LiDAR data [OTDS04, PV06, BH09, HDP09, RRP09, MEs10].

Some attempts have also been made to extract the road and its features from LiDAR data. Clode et al. [CKR04] segmented airborne LiDAR point cloud data into road and non-road objects using a hierarchical classification technique based on elevation and intensity information. The accuracy of their road segmentation approach was reduced due to the presence of car parks and private roads in their survey area. Hu et al. [HTH04] segmented LiDAR data into road and non-road areas based on elevation and intensity attributes. The Hough transformation was then applied to extract the

(46)

candi-date road stripes and parking areas. High resolution optical image data was also used to obtain road areas by extracting the concrete or asphalt pixels based on thresholding. Accuracy issues associated with the intermixing of road networks with parking areas were resolved using shape analysis and ve-hicle detection queues from the LiDAR and image data. Akel et al. [AKF+₀₅_] identified roads from airborne LiDAR data which were used for generalising the Digital Terrain Model(DTM). LiDAR data was segmented by applying a region growing approach on the basis of surface normal direction and height difference properties and then the extracted segments were classified into road and non-road objects based on a certain set of decision rules. Mum-taz et al. [MM09] identified buildings, trees and roads using a normalised Digital Surface Model (DSM) generated from LiDAR data and a Normalised Differential Vegetation Index (NDVI) estimated from high resolution aerial imagery. The resulting accuracy in road extraction was poor due to occlu-sions arising from buildings and tree shadows in the optical imagery. Oude Elberink et al. [OV09] developed an automated method for 3D modelling of highway infrastructure using airborne LiDAR data and 2D topographic map data. The road polygons were extracted from the topographic map data using a map based seed growing algorithm combined with a Hough transfor-mation. The LiDAR points were added to the corresponding road polygons using a LiDAR based seed growing algorithm. Subsequently, 3D reconstruc-tion was achieved by assigning the third dimension to the map polygons. Samadzadegan et al. [SBH09] used a multiple classifier system to classify the airborne LiDAR points into road and non-road objects using first pulse, last pulse, range and intensity attributes. Different combinations of LiDAR attribute layers were classified based on different features using maximum

(47)

likelihood and minimum distance methods. However, the optimum selection of features, type of classification technique and classifier fusion method were not conclusively addressed.

The majority of these road extraction methods attempt to delineate roads by distinguishing them from non-road objects but do make any attempt to extract the road edges. The road boundary is a fundamental feature, knowledge of which can provide precise estimation of other road features such as road markings and roughness. In the following sections, we review various methods developed for extracting road edges, markings and roughness from LiDAR data. We also investigate various methods based on snake curves which were developed for extracting road and urban features.

1.3.1 Road Edges

Road edges usually distinguish the road surface from kerbs in urban roads and from grass-soil in rural roads. Road edges with kerbs and grass-soil are shown in Figure 1.3. We do not define the edge between a road and a hard shoulder as the shoulders are used for emergency stopping or access. The hard shoulders can be extracted based on a similar approach used for road markings extraction as they both possess retro-reflective surface characteris-tics.

Road edges need to be correctly identified and extracted in order to obtain precise information about road geometry and physical road objects, . To date, little research has been focused on extracting precise road edges. Yuan et al. [YZC+₀₈_{] proposed an algorithm for extracting road surface from} terrestrial LiDAR data. The algorithm used a fuzzy clustering method to cluster LiDAR points. Straight lines were then fitted to the linearly clustered

(48)

Figure 1.3: Road edges: (a) kerbs in an urban road section and (b) grass-soil in a rural road section.

data using slope information for extracting the road surface area. Another approach for extracting the terrain surface from LiDAR point cloud data was formulated by Yoon et al. [YC09]. They calculated the slope and standard deviation characteristics from the LiDAR points and used these values to estimate the edges of the road. Vosselman et al. [VL09] developed a method for detecting kerbstones from airborne LiDAR data. The approach was based on the detection of small height jumps caused by the kerbstones in the LiDAR point cloud data. However, their extraction accuracy was affected by parked cars occluding the kerbstones. Zhang et al. [Zha10] proposed a method for detecting road edges in an urban environment using terrestrial LiDAR data. In their method, road edge points were identified based on elevation information. The identified 3D road edge points were then projected on a ground plane to estimate the road kerbs. Smadja et al. [SNG10] developed an algorithm for extracting roads from LiDAR data based on the detection of slope break points coupled with the RANdom SAmple Consensus (RANSAC)

(49)

algorithm [SNG10]. The extracted road boundaries were further processed to compute road curvature and road width information. McElhinney et al. [MKCM10] developed an algorithm for extracting road edges from terrestrial mobile LiDAR data. In the first stage of their algorithm, a set of lines were fitted to the road cross sections based on the navigation data and then LiDAR points within the vicinity of the lines were determined. In the second stage, these points were analysed along the Northing axis based on slope, intensity, pulse width and proximity to vehicle information in order to extract the road edges. The algorithm did not use the Easting values of LiDAR data to estimate the road edges.

Most of the methods reviewed have been developed for extracting road edges in an urban environment where algorithms rely on the existence of a sufficient height or slope difference between the road and kerb points for detecting road edges. Little or no research has been carried out to extract rural roads where the non-road surface comprises grass-soil and the edges are not as easily defined by slope changes alone. There is a need to develop a method that will provide an efficient and more accurate estimation of edges for different road types. Approaches developed to date make partial use of LiDAR data for extracting the road edges. The intensity and pulse width attributes from LiDAR data can be a useful source of information for ex-tracting these road edges. Their use in urban and rural road sections has yet to be thoroughly explored.

1.3.2 Road Markings

Road markings play an important role in reducing accident frequency and severity as they provide guidance and instruction to the road users for safe

(50)

and comfortable driving. They are intended to direct traffic by indicating the direction of travel, warn road users about specific obstacles or hazards and define the territorial limit for traffic flows [GPG+₀₇_{]. Road markings are} retro-reflective surfaces having an ability to reflect most of the incident light back to its originating source. These markings retain their visibility criteria in day and night. Examples of road markings are shown in Figure 1.4. Laser

Figure 1.4: Road markings: (a) continuous line, triangle and (b) arrow.

scanners usually record the reflectance of the illuminated road surface in the form of intensity data which can be used to distinguish road markings.

Smadja et al. [SNG10] extracted road markings by applying a threshold to intensity data acquired using terrestrial mobile laser scanning system. Vos-selman et al. [Vos09] recommended a normalisation of the intensity data prior to the threshold implementation or the use of a distance dependent threshold for extracting road markings from terrestrial LiDAR data. Jaakkola et al. [JHHK08] estimated road markings by first performing a radiometric correc-tion of the LiDAR intensity data using a second order curve fitting funccorrec-tion. Finally, road markings were estimated by applying a threshold and

(51)

morpho-logical filtering methods. Toth et al. [TPB08] used road pavement markings as ground control for assessing the positioning quality of aerial LiDAR data. The search window for finding the pavement markings in the LiDAR data was reduced by making use of the GPS survey data collected over the pave-ment. The pavement markings were extracted from LiDAR intensity values. Later, extracted pavement markings were compared with the GPS survey data to assess the quality of the LiDAR points. Chen et al. [CSW+₀₉_] devel-oped a method to extract lane markings from the terrestrial LiDAR intensity data. After detecting the road surface using the elevation information from the LiDAR data, lane markings were extracted by applying a threshold. The extracted lane markings were clustered using the Hough transform.

The majority of methods developed for extracting road markings are based on threshold applied to the LiDAR intensity values. The development of a robust threshold approach will provide a more precise extraction of road markings. A threshold applied to the intensity values often introduces noise, which needs to be reduced. The reflected intensity values depend upon the distance from the laser scanner to the illuminated surface, incidence angle of the laser pulse and surface characteristics. The intensity values need to be normalised in relation to these factors. A priori knowledge of the road boundaries and its surface will facilitate a more efficient extraction of road markings.

1.3.3 Road Roughness

The roughness of the road surface can be considered an important factor that influences safety condition for road users. It can be defined as the deviation of a road surface from a designed surface grade that may develop as a result

(52)

of road use, construction process or a combination of them [dFdS09]. Road roughness affects rolling resistance, ride quality, vehicle operating cost and safety of the road users [SK98]. Examples of roughness along the road sur-face are shown in Figure1.5. Several indices have been developed which are

Figure 1.5: Roughness along the road surface.

used to estimate roughness along a longitudinal profile of the road surface. These indices are computed as dynamic or geometrical values [dFdS09]. The dynamic indices such as International Roughness Index (IRI) provide contin-uous estimation of the roughness based on a model that simulates a dynamic response of a measuring vehicle along the road surface. The geometrical in-dices such as standard deviation of longitudinal roughness provide discrete estimation of the roughness in the form of standard deviation values of rela-tive elevation points measured along the road surface.

The IRI was developed by the World Bank in the 1980’s in response to a requirement for a reference scale for road roughness measurement [SGQ86]. This measure is used to provide a continuous estimation based on a model that applies a mathematical simulation of a standard vehicle moving along

(53)

the road surface profile at a certain speed. The model uses a Quarter Car Simulator (QCS), shown in Figure 1.6 [dFdS09]. The QCS consists of a

Figure 1.6: Quarter car simulator.

sprung mass that represents the vehicle body and an unsprung mass that represents the wheel and suspension. The sprung mass is connected to the unsprung mass with the suspension spring and damper. The unsprung mass is in a contact with the road surface using the wheel spring. During the simulation process, the QCS runs along the road surface profile at a constant speed V . The roughness along the road surface generates z_s0 and z_u0 vertical speeds in the sprung and unsprung mass respectively. The IRI value for a section of the road surface profile is estimated as

IRI = 1 L x/V Z 0 |z_s0 − z_u0|dt (1.1)

where L is a length of the road section in meters, x/V is a time taken by the model to travel a certain distance x and dt is a time increment. Thus,

(54)

the IRI is an accumulation of a vertical displacement and divided by the distance travelled by the vehicle that, in turn, provides the roughness scale. Its value is estimated in m/km or inch/mile units and ranges between 0 to 20 m/km. The 0 m/km value of IRI represents a perfectly smooth road surface, approximately 6 m/km value represents a moderate road roughness and 20 m/km value represents a bumpy unpaved road surface [Pat87]. One key advantage of using the IRI scale for the roughness measurement is its reliability as it facilitates both repeatability and stability of results with respect to time [SK98].

The standard deviation of longitudinal roughness provides discrete es-timation based on elevation points that are measured at a 1.5 m interval along the longitudinal profile of the road surface as shown in Figure 1.7

[dFdS09]. These elevation points along the longitudinal profile are measured

Figure 1.7: Measurement of elevation points at a 1.5 m interval along the longitudinal profile of the road surface.

using straight edge profilometers or laser profilers. The relative elevation di

for each point is computed as

di = hi−

1

2(hi−1+ hi+1) (1.2)

where hi, hi−1 and hi+1 are the current, previous and next measured

(55)

estimated as σ = v u u t n P i=1 (di− ¯di)2 n (1.3)

where ¯di is a mean of the di values and n is the number of points.

LiDAR data provides elevation values which have also been used for esti-mating road roughness. Pattnaik et al. [PHS03] estimated grade and cross-slope parameters of a road segment from LiDAR data. Road boundaries were delineated using multi-resolution orthophotos, GIS street database and a terrain model generated from LiDAR data. A centreline of the road was then determined using the estimated road boundaries. A plane was fitted to the LiDAR points using a linear regression model. Residuals for the grade and cross-slope were then estimated by finding a goodness of fit of the re-gression plane with the LiDAR points. Zhang et al. [ZF05] also presented a method for estimating road grade and banking from LiDAR data using a linear regression model, however road boundaries were extracted based on a priori knowledge of the road width instead of using a surface terrain model. Some other approaches have been developed for estimating roughness over soil, river bed and other terrain surfaces. Zhang et al. [ZR04] demonstrated prototype system for estimating a ground surface roughness information us-ing a combination of a camera and laser scannus-ing system. A multiscale variance method was used over different ground surfaces to characterise an elevation profile at different spatial scales. Hollaus et al. [HH10] investigated two approaches for estimating terrain roughness from full waveform airborne LiDAR data. In the first approach, an orthogonal regression plane was fitted to the LiDAR data and then the standard deviation values of the residual elevation points were calculated. In the second approach, the roughness

Road Features Extraction Using Terrestrial Mobile Laser Scanning System